What is the relationship between a president’s approval in the spring and his party’s House seat loss/gain in the upcoming midterm election? Charles Franklin discusses this question at Political Arithmetik. Approval in late April of around 36.5% (approximately where Bush was then) translates into a point prediction of around 45 seats lost. However, as Charles notes, and as one can learn through the application of interocular tools to his accompanying graph, the error on this estimate is quite large.
It is worth noting that 1994 is one of the biggest outliers, in that the Democratic party lost “too many” seats for the level of approval the public registered that spring for President Clinton. Of course, in 2006, with a swing of only 15 seats required to change control of the House, the Democrats do not need to over-shoot the estimate; they just need to avoid under-shooting it too much.
Quite apart from other factors that might lead to an over- or under-shooting of the estimated seat loss for a given approval level, there is no direct causal mechanism that links these two variables. The more proximate link is between approval and the president’s party’s vote share in the midterm election. Then a second stage translates any change in House votes into some level of change in House seats. One of our specialties here at F&V is votes-to-seats translation, and Charles was kind enough to share his data. So, let’s combined the approval and seat loss data with the national votes data and see what we get.
The first graph shows approval in late April/early May and change in House votes for the president’s party.
(Click the image to open a larger version in a new window.)
The 1950 election stands out as an outlier, but otherwise, there is clearly a trend in the expected direction: lower approval, more vote loss. Still, the scatter is hardly trivial. The 1950 election is an odd one. President Truman’s Democratic party lost seats (-29), even though it gained votes (+1.6). Truman’s approval was 41% at the beginning of May. Part of the reason for the Democrats’ vote gains was the high “other” votes total for the House in 1948, a year in which Strom Thurmond and Henry Wallace ran third-party campaigns for president. (Most analysts of US elections use the two-party vote, but I hate to throw data out; third parties are part of the picture in some elections. As a share of the two-party vote, Democrats indeed lost votes–very slightly.)
A statistical regression on these 14 data points shows–nothing. However, if 1950 is excluded, we do get a significant result. I will leave it to readers to decide if it is valid to exclude 1950. It is clearly is an outlier in a way that no other midterm election is, and it has the biggest movement in the past quarter century from one election to another in the the third-party vote. I think both of these factors make it valid to remove.
Simulating from the regression (without 1950), approval in April/May of 36.5% (Charles’s estimate for late April) predicts a vote change of -7.2, and a 95% confidence interval of -10.3 to -4.0. That is a rather big confidence interval, but the data are what the data are. Approval of 40% predicts vote change of -6.4 (-9 to -3.8); 50% predicts -4.3 (-5.7 to -2.8). The midterm loss phenomenon is a striking consistency in US politics (and, as my research has shown, that of other presidential systems, too): only with approval of 65% does the outer limit of the 95% confidence interval turn positive on the party’s vote change.
But control of the House depends only indirectly on the national vote share. Is there a systematic relationship between vote change and seat change in US House elections? I was prepared to find the answer to be not much, but actually there is. The graph below shows the relationship.
Again, 1950 looks a bit odd. So does 2002, one of only two US midterm elections since the Depression in which the president’s party gained seats. (The other was 1998, which looks fairly “normal” here, given that the Democrats in that year gained votes). In fact, it is worth noting that the way in which 2002 is “odd” is that the Republican party should have gained more seats. The party gained 3.3 percentage points over its 2000 House vote, yet gained only 8 seats. A regression on these data (again minus 1950) suggests the Republican party “should have” picked up around 20 seats in 2002, though the confidence interval is again large (4 to 36), and the 8 that they picked up is well within the 95% interval.*
So, if the Republican party loses 7.2 percentage points off its 2004 House vote share (the point prediction from 36.5% approval), it would be predicted to lose 49 seats (with a 95% confidence interval of -60 to -37). Even 37 is probably well beyond the worst fears of Republicans (though Rep. John Murtha (D-Penn.) predicted yesterday that his party would gain 40 to 50). The low-end estimate of the impact of 36.5% approval on the vote share (-4%) predicts a loss of 27 seats; even the most GOP-favorable end of the 95% confidence interval is 20 seats, which would be enough to shift control of the House.
In fact, any vote loss for Republicans of around 3.5% or more results in an estimate, even at the most GOP-favorable end of the confidence interval, of enough seats to shift control of the House. I am not in the prediction business, such a swing in votes strikes me as quite minimal: a 3.5% vote loss would be well within the margin of error of estimate even for a president with 50% approval.
There is one final analysis to make on these data. It is widely believed that House elections are less sensitive to national trends nowadays than they were decades ago. So we should consider whether there is a time factor to these estimates–both for the relationship between approval and votes change and for that between votes change and seats.
The short answer is that there is a substantial time effect on the first link, but none at all that is discernible on the second. In other words, the president’s party in the House is indeed more sheltered from assessments of the president than was once the case, as would be expected from the “personal vote” perspective or the incumbency advantage, or any models of voter choice that say that, for whatever reason, people tend to want different things from their president and their member of the congress nowadays, and more so than in the past. However, the absence of any time effect on the votes-seats relationship offers no support to claims that greater district homogeneity (from gerrymandering or any other factor) has made individual members more sheltered today than in the past. (And, yes, I am aware that this goes against much of the literature in my profession; I am just reporting what these data show, with due caveat for the substantial error of these regression estimates, as I have been noting throughout.)
To put this time analysis in real numbers (OK, estimates), in 1960 the party of a president with 50% approval in early May would expect a votes change of -5.65 (-7 to -3.9), while in 2004 such a party would have expected a vote change of only -2.1 (-4.3 to +0.2).
A president in 2004 with early May approval of 36.5% (such as the 2006 incumbent) could expect his party’s votes change to be -4.9 (-8.2 to -1.4). If the GOP can keep its vote losses to the more favorable end of that confidence range, it may be able to hang on to control of the House.
None of these statistical techniques allow us to predict what is going to happen in 2006, but they do tell us that the Republican party, if it can’t enjoy a substantial approval rally for its president, needs an unusually low linkage between that approval and its party vote share and/or an unusually low impact of the votes on the seats, in order to hang on. It’s a tall order.
*At zero change in votes, the estimate of seat change is indeed zero (actually -0.199), as it should be, and the 95% confidence interval is Â±10.